What I check in the last 15 minutes before merging an agent-generated PR

在合并 AI 生成的 PR 前，我最后 15 分钟会检查什么

The number one way that teams new to Agentic Engineering workflows accumulate long-term technical debt and scope drift is by merging agent PRs that look fine in the diff but break on edge cases: empty input, scope drift buried in a “refactor,” and tests that passed without testing anything. 对于刚接触“智能体工程”（Agentic Engineering）工作流的团队来说，积累长期技术债务和需求蔓延（scope drift）的首要原因，就是合并了那些在代码差异（diff）中看起来没问题，但在边缘情况下却会崩溃的 AI PR：例如空输入处理失效、隐藏在“重构”中的需求蔓延，以及那些看似通过但实际上什么都没测试的测试用例。

I wrote up the 15-minute pass I use before merge, including a worked example of what it catches on a synthetic agent PR. I just want to share in case others find it useful. 我总结了自己在合并前进行的 15 分钟检查流程，并附带了一个针对合成 AI PR 的实战案例，展示了该流程能发现的问题。我分享出来，希望能对大家有所帮助。